WO2008020557A1

WO2008020557A1 - Hand-written character recognizing method, hand-written character recognizing system, hand-written character recognizing program, and storage medium

Info

Publication number: WO2008020557A1
Application number: PCT/JP2007/065458
Authority: WO
Inventors: Shunji Mori; Tomohisa Matsushita
Original assignee: Kite Image Technologies Inc.
Priority date: 2006-08-14
Filing date: 2007-08-07
Publication date: 2008-02-21
Also published as: JP5352757B2; JP2008046825A

Abstract

The problem of symbolization is avoided, the structure is expressed in an analog way, and matching with a reference can be performed flexibly and simply. Each stroke of a hand-written character string is grasped in a parameter expression. The each stroke is approximated to a broken line. The broken line drawn by broken-line approximation is treated as a vector from the starting point to the end point. The angles between a reference axis and the broken lines are determined as a broken line angle sequence. The exterior angle sequence of the vertices of the obtained broken lines is determined. The sum of the exterior angles of the continuous same sign + or - of the exterior angle sequence is defined as a winding angle sequence. On the basis of the feature extraction by the determined sequences, nodes to be used as reference points are determined. A graph expression is given from the attributes of the positions of the nodes and the attributes of the sides between the nodes. Matching with a template having an open mask structure in which the starting point and the end point are not especially defined is so performed that character recognition is flexible and robust over the noise at the ends and deformation even if the character is one that holds normal up-down and left-right relations and even if rotation invariance is required.

Description

Specification

Handwriting character recognition method, handwriting character recognition system, handwriting character recognition program and storage medium

Technical field

The present invention relates to a handwritten character recognition method and handwritten character recognition system for performing online handwritten character recognition, a handwritten character recognition program for realizing the recognition method, and a storage medium storing the program.

Background art

[0002] A large number of character recognition systems have been proposed and put to practical use! /, The basic principle There are two positions, one is the position of structural analysis, and the other is the position of pattern matching. In the former, the recognition system is generally light, so the recognition system is generally light, that is, it is applied to an object with a fixed number of strokes, stroke order or constant, while the latter is both strokes. It has been applied to the case where the stroke order is free or close to it.

[0003] As a position of structural analysis, the Journal of the Institute of Electronics and Communication Engineers of Japan, 56-D, 5, pp. 312-319, "Online real-time recognition of handwritten numbers and Katakana characters" As described in JP-A-59- 1 31972, there is one called a basic stroke method, in which strokes are classified into simple strokes (four types) and combined strokes (seven types) and recognized by an identification automaton. Although it is simple, there have been problems with dictionary creation, cursing and abbreviations, and problems with their development.

[0004] The pattern matching method is roughly divided into two types. As described in the journal of the Institute of Electronics, Information and Communication Engineers, J63-D, 2, pp. 153-160, "On-line recognition of handwritten characters by point approximation of strokes", approximating strokes with a few points The motion direction of the brush at the end point is estimated with the feature point as the feature point, and they are also made into special focus to construct a feature vector. The dictionary is decomposed into strokes, and they have feature vectors as well, and the input vector is correlated with the feature vector prepared for each category, the distance is calculated for the corresponding dictionary, and the minimum distance is calculated. The given dictionary name is the recognized character name, and it is basically free for stroke order and stroke count. [0005] There is another pattern matching method, and the original article of the correspondence method of feature points is "Rubber String Matching method for handwriting recognition" described in the paper of PRL 74-20), In addition, JP-A-57-45679 and JP-A-8-24942 issued by the Japan Patent Office correspond to input character and feature point vectors of a dictionary according to DP (Dy n ic programming) method. This is the mainstream of online recognition of handwritten characters.

[0006] However, these methods are all complicated and satisfy the practical level in terms of performance sufficiently! /, It is the reality!

As flexible structural matching, research has been widely conducted in the field of scene recognition rather than character recognition. However, they are the arrangement of features in two dimensions and matching on a two-dimensional graph that generally expresses their relationship. There is a great deal of research on this.

On the other hand, in online character recognition, an image seen by a human being is two-dimensional force, which is strictly on the time axis, and is one-dimensional. That is, it can be expressed as a simple one-dimensional linear graph. This point of view dramatically simplifies the problem. Moreover, due to the winding angle, natural clipping candidate points are arranged in a negative manner on the time axis. The winding angle and the linear (one-dimensional) graph are the core of the present invention.

[0008] The present invention was made in view of force, and basically belongs to the structural analysis method described above, but overcomes the problems so far and is flexible structural analysis It provides the basis for the method, and therefore aims to avoid symbolization problems, represent structures analogically, and perform flexible and simple matching with standards.

Disclosure of the invention

In the present invention, in recognizing online handwritten characters, input handwritten character strings are captured for each stroke by parameter expression, and each line is subjected to polygonal line approximation, and each polygonal line is approximated by polygonal line. The angle between the reference axis and each polygonal line is determined as a polygonal line series as the vector extending from the start point to the end point, and the external angle series of each vertex of the obtained polygonal line is determined. The sum of the external angles of the same code where the codes are continuous is taken as a winding angle series, and based on the feature extraction by each series obtained, the point which is a reference point is no The graph representation is given by the attribute as the point of the node and the attribute as the edge between the nodes, and the start point and the end point are not particularly defined! / The character by the matching with the template by the open mask configuration It is to recognize.

According to the present invention, when recognizing on-line handwritten characters, it is flexible and noise at the end, even for characters that maintain normal upper / lower and left / right relationships, and also when rotation invariance is required. In the same way as the recognition method of the character written in isolation, it makes strong recognition in the continuous character string and makes strong recognition to the character and transformation. it can.

Brief description of the drawings

FIG. 1 is a block diagram showing an example of a system according to an embodiment of the present invention.

FIG. 2 is a flow chart showing an example of processing of the entire character recognition according to an embodiment of the present invention.

FIG. 3 is a flowchart showing a detailed example of character recognition processing according to the embodiment of the present invention.

FIG. 4 is an explanatory drawing showing an example of polygonal line approximation according to an embodiment of the present invention.

FIG. 5 is an explanatory diagram showing an example of a node according to an embodiment of the present invention.

FIG. 6 is an explanatory drawing showing an example of polygonal line approximation according to an embodiment of the present invention.

FIG. 7 is an explanatory drawing showing an example of polygonal line approximation according to an embodiment of the present invention.

FIG. 8 is an explanatory drawing showing an example of polygonal line approximation according to an embodiment of the present invention.

FIG. 9 is an explanatory drawing showing an example of the relationship between nodes and sides according to an embodiment of the present invention.

FIG. 10 is an explanatory view showing a determination example according to an embodiment of the present invention.

BEST MODE FOR CARRYING OUT THE INVENTION

Hereinafter, an embodiment of the present invention will be described with reference to the attached drawings.

The example of the present embodiment is applied to a system that performs online handwritten character recognition, and FIG. 1 shows an example of a configuration in which each processing unit has a hardware configuration. As shown in FIG. 1, the handwritten character recognition of this example is programmed in a general-purpose arithmetic processing unit such as a personal computer device or a general-purpose arithmetic processing unit in which each processing unit is executed by a common arithmetic processing unit. To implement the same handwriting recognition. Well, yes.

In the following description, the concepts required for handwriting recognition in this example are defined in terms as shown in Table 1 below.

[0014] [Table 1] Definition of terms.

To describe the configuration shown in FIG. 1, writing on the paper 1 with the pen 2 detects the pen stroke (la) on the paper 1 on the pen 2 side. The pen is detected by, for example, a pen 4/1

It does with the camera built in 2. Alternatively, the movement of the force pen 2 itself such as an acceleration sensor may be detected. Furthermore, even if the side of the paper 1 that is not detected by the pen is configured with any force panel, it is possible to detect the handwriting electrically. In any case, in the case of this example, since it is online handwriting character recognition, it is configured to be able to judge the deterioration of the handwriting over time.

[0016] The handwriting data detected by these processes are sent to the input processing unit 3 to output character information ^{: a} second input is performed. The input data is sent to the polygonal line approximation unit 4, the eyebrow extraction unit 5, the identification unit 6, and the identification result output unit 7, and corresponding processing is performed in each of the processing units. The identification result output unit 7 performs output processing such as display of the identified characters and 'output of the character code of the identified I. Marking or printing of the identification character may be performed based on the identified character code.

The flow chart of FIG. 2 shows an example of the entire processing of the character recognition of this embodiment. Below

Replacement paper (Rule 26) As described below with reference to FIG. 2, the character / graphic pattern input from the input processing unit 3 (step S11) is subjected to polygonal line approximation by the polygonal line approximation unit 4 (step S12). From this approximation, the input pattern is expressed as a vector having the length, the direction angle, and the difference in the direction angle of the adjacent line as an element when viewing each line as a vector (step S13). Also, from the vector expression of the difference of the direction angle, the sum of the difference of the same sign is obtained, and as one element including the code, the vector expression named as the winding angle is obtained here. Next, features are extracted according to the situation from the polygonal line approximation representation in the feature extraction unit 5 (step S14), and a one-dimensional linear graph representation based on the extraction results of the features is given (step S15). Character recognition is performed by matching the expression with a template based on an open mask configuration that does not particularly define the start point and the end point (step S16), and the character recognition result is output (step S17).

Here, an example of the details of the character recognition process in step S 16 will be described with reference to the flowchart of FIG. 3. First, it is checked that the arrangement of each node shape of the input graph expression matches (step S21). Then, the matching of the attribute as the node point is checked (step S22), and then the matching of the attribute of the edge between the nodes is checked (step S23). Finally, the matching of other attributes such as the presence or absence of the intersection point and the distance relationship between nodes is checked (step S24) and identified. If all the results of these checks match, the recognition result is OK (step S25), and if there is not even one match, it is excluded (step S26).

The recognition method in the present embodiment basically belongs to the structural analysis method described above, but it overcomes the problems so far and provides a basis for a flexible structural analysis method. It is a thing. Therefore, it avoids the problem of symbolization, expresses the structure in an analog manner, and performs flexible and simple matching with the standard. In addition, structural analysis is performed, so that the subject can inevitably be described properly, and the correspondence between cause and effect is clear from the human vision. Therefore, it is possible to evaluate the shape of objects such as letters, set the correct rejection range, and provide a recognition system with more human-like ability.

Up to this point is basically the same as the structural analysis method previously proposed by the inventor of the present application.

In addition, this method is very convenient for efficient character recognition in continuous character strings (eg cursive letters etc.). In the present invention, the winding angle makes it possible to Focusing on the fact that the extraction candidate points are positively aligned in the negative, the input character string is one-dimensionally linear using the attributes of the candidate points (nodes) and the edge between the nodes (edges) A graph representation is given, segmentation is performed simultaneously with recognition ("segmentation recognition"), and segmentation can be performed in the same way as the recognition method for isolated characters written before preprocessing. is there.

[0021] In explaining the effects of the present invention, it is easy to enter from recognition of continuous character strings, so we will start with the explanation of the specific digit string recognition method shown in FIG. This is "888". These letters are connected. Therefore, in conventional character recognition, these characters had to be cut out and sent to the recognition system. This process is called "pre-cut process". However, in this method, this “pre-cutout process” is unnecessary, and recognition and cut-out are performed simultaneously. This is explained below.

[0022] The image in Fig. 4 is a force written from the left. The basic representation of these is the length of each polygonal line to be connected, the polygonal line angle, the outer corner composed of the adjacent polygonal line corners, and the same sign. The length series, the angle series, the outer angle series, and the winding angle series that have the winding angle as an element, which is the sum of the outside angles of the issue.

Therefore, these are used as raw materials and converted into a one-dimensional graph representation suitable for “cutout recognition”. Here, the graph is the simplest linear graph. The following is an example of Fig.5, which is the first "8" part of Fig.4. Here, a start point, an end point, a point giving a winding angle of an integral multiple of 90 degrees (± n × 90 degrees long point), a point where the sign of the winding angle changes is a “node”, and that point is used as a node attribute. Maximum value of the absolute value of the outer angle in the group of broken lines between the nodes as an attribute of the outer angle (△) of the area, the winding angle value (Θ) of the winding angle area to which the node belongs, and the side (edge) between nodes Use | Δ |) to give a graph representation.

[0024] Here, 0 0 [s] 1 1 [1] [2] [1 +] 3 [[]] [4] [+], 5 [[5] [+]

], ················································································································································································ · It starts from く 0 [[s] and ends at く 28 [[e]. [s] and [e] are start and end symbols, respectively. 1) [-] is the first winding angle-90 degree point in the outer angle (Δ) series. Since the winding angle is discrete linear interpolation, it can generally be obtained in analog form as an α degree long point. 1 一 [-] "one" is a left angle mark Express the issue.

Graph representation of Figure 5

0] [s] 0 (326, 121), Θ = -172.83

I

I number of polylines (0 to 4) = 4, length = 0.07, maximum I Δ I = | -24.52 |

I

1 [1 4 4 (156, 121), Θ = -172.83, Δ = -33.26

I

I number of polylines (4 to 6) = 2, length = 0.09, maximum I Δ I = | -72.64 |

I

2) [1 +] 6 (347, 279), Θ (-) = -172.83, Θ (+) = 304.42, Δ (-) = -72.64 Δ (+) = 21.27

I

I number of polylines (6 to 8) = 2, length = 0.02, maximum | Δ | = | 45.06 |

I

3> [+] 8 (365, 346), Θ = 304.42, Δ = 49.31

I

I number of polylines (8 to 12) = 4, length = 0.07, maximum I Δ | = | 30.07 |

I

4> [+] 12 (195, 329), Θ = 304.42, Δ = 20.39

I

I number of polylines (12 to 15) = 3, length = 0.03, maximum I Δ | = | 35.79 |

I

5) [+] 15 (194, 258), Θ = 304.42, Δ = 38.45

I

I number of polylines (15 to 16) = 1, length = 0.05

I

6) [+-] 16 (330, 205), Θ (−) = 304.42, Θ (+) = − 56.33, Δ (−) = 38.45, Δ (+) -56.33

<28> [e]

[0026] [1] [-], that is, the "one-90-degree point" is slightly beyond the fifth vertex. Therefore, as an attribute of the point of this node, the outer angle (Δ) value of the winding angle series (Θ = -172.83) to which it belongs and the closest previous vertex (in this case, the fifth vertex), Δ Give = -33.26. In addition, 2> [-+] is important to indicate the boundaries of winding angles. This is the point where the winding angle (one) changes to the winding angle (+). Boundaries of winding angles share one broken line at points. The tip of the shared polygonal line in the positive direction of the time axis is taken as a node. The attributes of this point are Θ (−) = − 172.83, Θ (+) = 304.42, Δ (−) = − 72.64, Δ (+) = 2.27. That is, the winding angle Θ (−) = − 172.83 before temporal change, the winding angle Θ (+) = 304.42 after temporal change, the external angle (Δ) value at both ends of the common polygonal line, Δ (−) = _7 2.64, Δ (+) = 21.27.

[0027] From 2 [[-+] onwards, the polygonal line series enters the area of winding angle (+). This is a 90 ° long point at the beginning 3> [+]. The attributes are Θ = 304.42, Δ = 49.31.く 4 [[+] is a long point of 180 degrees. 5) [+] is a 270 degree long point. That is, at this node 5>, the winding angle is over 270 degrees. Next is 6> [+-], which is the point where the winding angle changes from + to one.

Next, as an attribute of an edge connecting these nodes, the number of broken lines (0 to 4) = 4, length = 0.07, maximum |

Take the maximum value of the number of broken lines, length, | Δ | between these nodes according to Δ | = | −24 · 52 |. These properties are rotation invariant.

The above is the graphical representation of the input, and on the other hand, matching with the template by the mask configuration is performed. An example of the mask configuration in the case of "8" is shown below.

[0030] <Example of "8" Mask>

Condition 1: * = one + (first key node)

Condition 2: -200 <Θ (*-) <100 & 200 <Θ (* +) <360 (Note: "Θ (*-)" is the first "" winding angle, "Θ (* +)" is the succeeding "+" winding angle) Condition 3: -100 mm Δ (* _) K -20 & 10 Δ (* +) 100

(Note: “Δ (* −)” is the outer angle value of the “one” winding angle boundary, “Δ (* +)” is the outer angle value of the succeeding “+” winding angle boundary, ie, the boundary line segment Side on the both sides, + outside angle value) Condition 4: (*-, * +), Cross: nxm: ne ^-) ~ (* +)

(Note: There is an intersection Cross between the two ends of the boundary line)

Condition 5: * + 1 = +-(second key node)

Condition 6: 200 Θ (* + 1 +) 360 360 &-200 Θ)-10

(Note: "/ Θ (* + 1 +)" is + winding angle, "Θ (* + 1-)" is next winding angle

Condition 7: 10 times Δ (* + 1 +) 100 +-100 Δ (* + 1-)-20

(Note: “Δ (* + 1 +)” is the external angle value of the + wrap angle boundary, and “Δ (* + 1−)” is the external angle value of the next wrap angle boundary.)

Condition 8: (* + 1_, * + 1 +), Cross: nxm: mE (* + l +) to (* + l—)

(Note: Cross exists between both ends of boundary line) Character [8] = Condition 1 & Condition 2 & Condition 3 & Condition 4 & Condition 5 & Condition 6 & Condition 7 & Condition 8

[0031] This mask is very simple, and there are two nodes ("* +" and "* + 1 = +-"), respectively, and the boundaries from-to +, + to-only. Yes, “± n x 90 degree long point” is not used. Here, the meaning of "*" in "* +" indicates that this node is the key of this mask, and indicates that there is always a node that changes from winding force S "minus" to "plus". . "* + 1 = +-" is the next required node, and this time, raise the node whose winding angle changes from "plus" to "minus" this time!

In addition, as the attributes of the first key node in condition 2, condition 3, and as the attributes of the second key node in condition 6, condition 7, the value of the winding angle, at both ends of the boundary line segment Detailed conditions are given to the outside angle Δ of, and the structure is tightened. Also, intersection point information is used as an important edge property. That is Condition 4 and Condition 8. me (* + l +) to (* + l −) require that an intersection be present between the boundary lines at the next winding angle change point. Cross the intersection with Cross: nxm expressing. The n and m are the numbers of intersecting polygonal lines, and ne (*-) to (* +) indicate that the intersecting polygonal lines coincide with the winding angle boundary line segments. Of course, the intersection is a rotation invariant feature.

[0033] The above is a very simple mask. This is matched with the input graph representation, and the node range in which this type fits is searched, and if it exists, it is assumed that "8" exists there. It should be noted that the above mask does not use the [s] and [e] nodes. Thus, both ends are open. For this reason, in the flow of time, if there is a place (range) that applies to this mask, it means that there is "8", and "pre-cutout processing" regardless of the place at all. Is unnecessary.

In fact, in this mask, three “8” s are recognized in this mask. Also, in FIG. 6, connected "8" s, which have completely different inclinations, are respectively recognized. Also, in Fig. 7, the force at which two "8" s are connected, which are significantly different in size, is also correctly recognized as "8". As a matter of course, it is possible to recognize a single character as well by combining similar masks and input graphic expressions, regardless of connected characters.

Here, some points to note will be described. That is, the mask mentioned above does not include the characteristic of full distance or length. Therefore, regardless of the length of the connection, the shape was fitted everywhere, and the cut-out recognition was made. However, on the other hand, for example, as in the case of Fig. 7, the information of two "8" sizes is lost, and there is a doubt that the difference in size can not be understood as seen by humans. However, this can actually be easily determined.

This input representation has position coordinates of each node. Therefore, for example, in the case of "8", the node returned 180 degrees from the "*" node which is the mask key, and in FIG. 7, this is almost an <s> node. Find the Euclidean distance between the positions of

[0037] Specifically, it is the square root of (286-38) 2 + (562-449) 2 for large "8" = 273 for small "8" (174-55) 2 + (: 625- 611) The square root of 2 = 120, and the ratio of the two is about 0.43. Although it is somewhat complicated, in the case of FIG. 2 also, the vectors of the major axis and minor axis of both “8” can be determined, and inclination information can also be known. Thus, the coordinate values of the node can be effectively used to obtain geometric metrics such as length, angle, direction, etc. is there.

In the example of the mask configuration of “8” described above, the force described in the case of a simple configuration in which the ± n × 90 degree node is not added to the condition is used here to explain the essence of the present invention. The effectiveness will be described using an example. This is particularly effective when it is round, the typical being a circle.

[0039] This is actually a symbol that is often used as "o". The challenge is to make a mask that holds the shape of this circle and that is resistant to considerable deformation.

This mask is made as follows.

Condition 1: * = +90

Condition 2: 350 Θ (*) 600 600 & 0 Δ Δ (*) 95 95

Condition 3: (*. * + 1); 0.1 length 0.35 & 0≤ | Δ |

Condition 4: * + 1 = +180

Condition 5: Θ (* + 1) = Θ (*) & 0 く Δ (*) 95 95

Condition 6: (* + 1 · * + 2); 0 · 1 <length 0.35 & 0≤ | Δ |

Condition 7: * +2 = +270

Condition 8: Θ (* + 2) = Θ (*) & 0 く Δ (*) く 95

Condition 9: (* + 2 · * + 3); 0 · 1 <Length 0.35 & 0 ≤ | Δ | <95

Condition 10: * +2 = +360

Condition 11: Θ (* + 3) = Θ (*) & 0 Δ Δ (*) 95 95

Condition 12: (* + 3. * 4); 0 · 0 <Length 0.35 & 0 ≤ | Δ | 95 95 Sign "〇" = condition 1 & condition 2 & condition 3 & condition 4 & condition 5 & condition 6 & condition Condition 7 & Condition 8 & Condition 9 & Condition 10 & Condition 11 & Condition 12

There are 4 nodes ("* = + 90", "* + 1 = + 180", "* + 2 = + 270", "* + 3 = + 3 60") of this mask, ± Χ Χ 90 degree node Existence (in this case + Χ 度 90 ° node) Force S, roughly defines this form. That is, in this case, n is 1 to 4 and the winding angle is 360 in units of 90 degrees. It is requested to wind up to a degree. On the other hand, as for the attribute of the node, the sharp increase of the angle is suppressed by 0 and Δ (*) and 95 in the vicinity of + n X 90 degree long point. Also, as an attribute of the edge, it is required that the length between nodes be within a certain range. This is "0.1 long and 0.35". It also stipulates that there will be no sharp increase in the angle between them. This is "0 rd | Δ | s 95".

[0041] This mask is sufficiently configured to be resistant to noise! /, And both ends are open //.

Therefore, as shown in FIG. 8, “〇” is recognized even when there is considerable external noise. If each length is normalized relative to the length as described above, for example, with the distance between the 180 ° long point and the 360 ° long point, this can be an arbitrary number of concatenated character sequences. Is also applicable.

Here, we do not need the usual rotation invariance! /, And the graph representation for character recognition! In such a case, it is more intuitive and easier to use as a node the unevenness based on the horizontal and vertical orthogonal coordinate system rather than using the ± Χ Χ 90 degree node. Please see the illustration of Figure 9 as an image. Here, 〇 is a node, and an arrow → is a side.

Here, the upper and lower irregularities are indicated by u and ΓΊ, and the irregularities viewed from the left and right are indicated by 、 and c. Besides, in fact, the upper and lower asperities and the left and right asperities simultaneously exist. For example, in the case where it is in the upper right, it is indicated as n + co where n and one overlap.

[0044] The attributes of the nodes are as described above. As the attributes of the force side, the average direction angle of the polygonal line group between the nodes is added. This is a very important feature in concavo-convex node expression. Others There is a directional variance of the polyline group indicating the degree of bending. Next important is the distance ratio between nodes. For example, referring to Figure 9, the start point must be somewhat above the end point. These can be quite large when calculated mechanically, but in practice they can not be good numbers if only the point is kept down. Here, for example, the start point <s> node Y axis value is subtracted from the Y axis value of node 5>, and the value obtained by dividing by the ratio of the length of the whole character is obtained.

Hereinafter, an example of the mask of “8” by this concavo-convex graph expression will be shown.

Condition 1: * -1 = s / any Condition 2: (*, * -l), 110 weighted average angle 170

Condition 3: * = n

Condition 4: -420 degrees Θ (*)--160

Condition 5: -110 square Δ (*) square -10

Condition 6: * + 1 = 匚

Condition 7: Θ (* + 1) = Θ (*)

Condition 8: -110 ° Δ (* + 1) ° -10

Condition 9: (* + 1, * + 2), -80 weighted average angle-30

Condition 10: (* + 1, * + 2), length> 0.15

Condition 11: * + 2 =

Condition 12: 250 Θ (* + 2) 500 500

Condition 13: 20 times Δ (* + 2) 100

Condition 14: * + 3 = U

Condition 15: Θ (* + 3) = Θ (* + 2)

Condition 16: 10 times Δ (* + 3) 120

Condition 17: * + 4 = c

Condition 18: Θ (* + 4) = Θ (* + 2)

Condition 19: 10 times Δ (* + 4) 100

Condition 20: (* + 4, * + 5), 20 weighted average angle 75

^ {ψ 21: (* + 4, * + 5), Cross: nxm: ne (* + 1) ~ (* + 2), m ≡ (* + 4)-(* + 5)

Condition 22: * + 5 = e / any

Condition 23: ((* + 1, * + 2) Cross Y value-(*) Y value) / Height 0 · 20 <Υ <0 · 70〃 The intersection is located approximately at the center in height. Note the characters Letter '8' = condition 1 & condition 2 & condition 3 & condition 4 & condition 5 & condition 6 & condition 7 & condition 8 & condition 9 & condition 10 & condition 11 & condition 12 & condition 13 & condition 14 & condition 14 & condition 15 & Condition 16 & Condition 1 7 & Condition 18 & Condition 19 & Condition 20 & Condition 21 & Condition 22 & Condition 23

In this case, no particular inter-node distance ratio is required. Instead, the genus of the side regarding the intersection Condition 23 which is sex plays the role. Thus, the inter-node distance ratio is actually much less than the number produced in combination. That is because it is restricted within silence by the restrictions of wrap angle, outside angle, and line angle. Also, the length restriction in the upper mask only limits the length of the polyline, which goes from left to right. In fact, because of this, very, strong noise on the edge, it becomes a mask!

Note that this is not a cutout recognition mask for continuous character series. That is because we put a limit on this length. In this respect, if this length is normalized by the Y-axis value between the node * = ΓΊ and the node * + 3 = U, it becomes a cutout recognition mask for a continuous character series.

[0048] With this method, even when extreme noise occurs at the end or extreme deformation, as long as it has the core "8" shape, it has nothing to do with the surrounding situation. The correct recognition is

The cut-out recognition method described above necessarily has the possibility to give multiple answers.

A specific example of this is shown in FIG. Figure 10 is written with the intention of “8”. However, “6” is hidden in this figure. If you change the word, the character form "6" is "cut out and recognized" from the original figure. Therefore, this is an inevitable result.

[0050] Therefore, the intended mechanism for correctly outputting "8" is required. Intuitively speaking, this means that the complex shape is to be prioritized. Forces To be stated quantitatively, the length of the line matching the "6" mask and the mask of the "8" match Naturally, the latter becomes longer if the length of the broken line is compared.

From the following example, “8” is longer than “6” by the length of 0.27. Therefore, "6" is quantitatively buried in the image of "8", and "8" has a matching portion longer than "6" by 0.27 and is determined as "8". In addition to this, measures of matching such as the number of matching nodes and the number of broken lines can be considered.

Note that, as described at the beginning of the description of the embodiment, the handwritten character recognition of the present invention is substantially the same handwritten character recognition that is not limited to the processing configuration shown in FIG. To perform recognition processing with various device and system configurations. It is possible. For example, the handwritten character recognition of the present invention is converted into a program (software) to make a general-purpose personal computer.

It may be implemented on a computer device. The handwritten character recognition program can be stored in various storage media and distributed.

In this case, the character recognition may be performed on the off-line characters by the force S for the on-line characters, appropriate thinning, or contour tracing.

Furthermore, in the above-described embodiment, although the case of mainly recognizing numbers is taken as an example, handwritten character recognition of the present invention can be basically applied to characters and symbols of any language. Description

[0054] 1 ... paper, la ... stroke, 2 ... pen, 3 ... input section, 4 ... polygonal line approximating unit, 5 ... preprocessing unit, 6 ... characteristic extraction unit, 7 · · · · · · · Identification unit, 8 · · · output unit

Claims

The scope of the claims

[1] Online handwriting recognition method /!

Capture the input handwritten character string for each stroke by parameter expression, and perform line approximation for each stroke,

The angle between the reference axis and each polygonal line is determined as a polygonal angle series, with each polygonal line approximated by the polygonal line as a vector extending from the start point to the end point.

Find the outer corner series of each vertex of the obtained polygonal line,

Let the sum of the external angles of the same sign where the same sign of plus or minus of the external angle series continues is the winding angle series,

Based on the feature extraction based on each of the obtained series, a node as a reference point is determined, and a graph representation is given by the attribute as the point of the node and the attribute as an edge between the nodes. By matching with the template by the open mask configuration which is not specified, it is flexible and strong against the noise and the deformation of the edge, even for the character that holds the normal upper and lower, left and right relationship, and when the rotation invariance is required. A character recognition method characterized by robust recognition.

[2] In the character recognition method according to claim 1,

Start point, end point, point giving winding angle of integer multiple of 90 degrees, point where winding angle changes as node, give graph representation by attribute of node as point and attribute of edge between nodes, specify start point, end point It features rotation-invariant recognition means by matching the template with a non-open mask configuration.

Character recognition method.

[3] In the character recognition method according to claim 1,

In order to recognize ordinary character shapes that do not require rotation invariance, let the upper and lower asperities and left and right concave and convex nodes be nodes, and give a graph representation by the attribute as the point of the node and the attribute as the edge between the nodes. It does not specify the start point and the end point in particular! / Character recognition is performed by matching with the template by the open mask configuration

Character recognition method.

[4] In the character recognition method according to claim 1,

In the case where a plurality of characters or symbols are in an inclusion relation, it is characterized in that the matching portion preferentially gives priority to the long characters or symbols.

Character recognition method.

[5] In the character recognition method according to claim 1,

A character recognition method characterized in that continuous character string characters are recognized and cut out simultaneously with recognition.

[6] A handwriting recognition system that recognizes online handwriting!

Input means for inputting handwritten characters online;

A broken line approximation means for capturing input handwritten character strings as parameter expressions for each stroke and performing polygonal line approximation for each stroke;

Each broken line approximated by the broken line approximation means is a vector that is calculated from the start point to the end point, and the angle formed by the reference axis and each broken line is obtained as a broken line angle series, and each vertex of the obtained broken line A processing means for obtaining an outer angle sequence, and taking the sum of the outer angles of the same sign in which plus or minus same symbols of the outer angle sequence are continuous as a winding angle sequence;

Based on feature extraction by each series determined by the processing means, a node which is a reference point is determined, and a graph is represented by an attribute as the node point and an attribute as an edge between the nodes.

If you can find rotation invariance even for characters that maintain normal upper / lower and left / right relationships by matching with a template with an open mask configuration. Also have steps to perform robust recognition that is flexible and resistant to noise and deformation at the edges.

Character recognition system.

[7] Write a handwriting recognition program that recognizes handwriting online,

Capturing handwritten input character strings for each stroke by parameter expression, and performing polygonal line approximation for each stroke;

Obtaining the angle between the reference axis and each broken line as a broken line angle series, using the broken line-approximated broken lines as vectors extending from the start point to the end point; Determining an outer angle sequence of each vertex of the obtained polygonal line;

Setting the sum of the external angles of the same sign where positive or negative same signs of the external angle series are continuous as a winding angle series;

Based on the feature extraction based on each of the obtained series, a node as a reference point is determined, and a graph representation is given by the attribute as the point of the node and the attribute as an edge between the nodes. By matching with a template with an open mask configuration that is not specified, it is flexible even for characters that maintain normal upper and lower, left and right relationships, and when rotation invariance is required. It is characterized in that it has a step of making strong and robust recognition.

Handwritten character recognition program.

[8] A storage medium capable of online handwriting character recognition by mounting a stored program on a predetermined arithmetic processing unit,

As a program stored in a storage medium,

Obtaining the angle between the reference axis and each broken line as a broken line angle series, using the broken line-approximated broken lines as vectors extending from the start point to the end point;

Determining an outer angle sequence of each vertex of the obtained polygonal line;

Based on the feature extraction based on each of the obtained series, a node as a reference point is determined, and a graph representation is given by the attribute as the point of the node and the attribute as an edge between the nodes. It is flexible and resistant to edge noise and deformation, even for characters that maintain normal upper and lower, left and right relationships, and when rotation invariance is required, by matching with a template with an open mask configuration that is not specified. And a step of performing robust recognition. Storage medium.